A Call for Europe to Wag the Dog

We don't need the capital to win the frontier. We only need to understand where the food enters the animal — and the nerve to let the tail wag the dog.

There is a quiet war being fought at the frontier of AI, and like all good wars, it is being lost — slowly, expensively — by precisely the people who declared it.

The premise is reasonable. Every lab wants the best model, wants to knock the others off the leaderboard, wants to hold the edge just long enough to put a price on it. Fair. That is competition, and competition, we are assured, is good for me — the user.

Except the edge arrives wrapped in a quota.

So my working day looks like this. I open a problem in ChatGPT and get it moving. The meter taps me on the shoulder. I carry the half-finished thing over to Gemini and push it further. Gemini tires too. I hand it to Claude to tidy up, take a quick detour through DeepSeek for a second opinion, and by the time I've gone round the table the first quota has quietly refilled — and the loop begins again.

Notice what I am actually doing. I am not using four products. I am a courier. I take the output of System A, walk it across the street, and feed it into System B — then run it back the other way. Every lab's best work passes through my clipboard on its way into a rival's context window. I am a one-person data-laundering operation, and I run it for free, all day, cheerfully.

Which brings us, inevitably, to the lawyers. Somewhere upstream, Company A is drafting a stern letter accusing Company B of training on its outputs. Distillation! Theft! The crown jewels, siphoned! And they are not entirely wrong — they are just hunting the wrong tunnel. They picture a rival scraper boring into the vault. The real tunnel is the clipboard. The smuggler is me. The route is normal usage. And the incentive was written, personally, by whoever set the quota that asked me to leave the moment the meter ran out.

Sorry. But not sorry. ;)

So let me offer a redefinition, no charge. A quota is not the amount of work you'll do on Platform A. That number is a fiction. A quota is the amount of Platform A's unfinished work the user will export to Platform B to get the job done. You didn't cap my usage. You set my export volume.

Now turn the tube around.

From the lab's own chair the same number means something stranger. A generous quota doesn't only decide how much of me you consume — it decides how much of my competitors you're carrying when you walk in the door. The more I give you, the less rival data I receive, because you stay and finish here. The stingier I am, the more I leak — you take my half-chewed work and finish it on somebody else's screen. The generous model becomes the sink: the bottom of the watershed, where every model's prechewed, multi-pass, lovingly-revised work eventually pools. So the miser pays twice — he gives away his own half-digested work upstream, and refuses delivery of everyone else's downstream.

And I can prove it from this week. I pass Claude's results into ChatGPT far more often than the reverse — not because either is better, but because ChatGPT lets me keep going, so that's where the finishing happens and that's where Claude's work comes to rest. This very morning, the moment Fable declined to carry a task any further, DeepSeek was delighted to pick up the prechewed remains and run with them. The data did not flow toward the best model. It flowed downhill — from the throttled to the permissive — and gathered wherever I was simply allowed to stay.

Now, before some strategist reaches for the obvious rebuttal — yes. There is a compute cost. Of course there is. Frontier AI is not made of slogans; it runs on chips, power, engineers, data centres, money and time, and no clipboard on earth conjures a GPU. So build the datacentres. Build all of it.

But curated data lowers the cost of intelligence. A model fed garbage must burn compute discovering structure from scratch. A model fed real tasks, strong drafts, human corrections and multi-model refinements inherits structure already half-paid-for by someone else. What I'm ferrying between these labs is not internet sludge — it's live, high-intent, multilingual, already-refined human work, exactly the material every lab swears it cannot get enough of. I am not a replacement for their compute. I am a multiplier on it. I improve them by abandoning them. The act of leaving is the contribution.

There's one honest catch, and I'll say it before the comment section does: this only bites where the product terms let it. Paste into a model that never learns from the chat and you've fed nothing but its ego. So the lever isn't only where users finish — it's also what the terms permit. Fix the terms, and the tail grows teeth.

Which is where Europe ought to be paying attention, and mostly isn't. We keep talking as though sovereignty can only arrive by billion-euro datacentre, sovereign-cloud tender, committee, alliance and heroic capital. Fine — do all of it. But don't ignore the free lever sitting in three hundred million hands.

The Americans have the capital. The Chinese have the scale. Europe has something neither can fake overnight: hundreds of millions of educated, multilingual, regulation-aware users sitting on industrial, legal, scientific, administrative, agricultural and commercial workflows — the messy, specific, real European weirdness that exists nowhere else. That is not "just usage." That is the data layer.

So here is the doctrine, and it costs nothing. Start wherever you like. Use the best model for the first pass — the American one, the Chinese one, the closed one, the clever one, the cheap one. But finish European. Let the final version pool here. Let the corrections, the hard cases, the bureaucratic forms, the procurement drafts and the engineering notes land here. Paste into Mistral. Work it, break it, correct it, feed it — not out of charity, but out of strategy, and where the terms allow, out of appetite.

If frontier AI is a dog, Europe does not need to own the biggest head in the room. It needs the tail that feeds it — and the tail, dear fellow Europeans, is us. No gigawatt datacentre required to start. Only a clipboard, and a small, stubborn patriotism.

Sorry.

(Not sorry.)